| 摘要: | 癌症是一種由基因、轉錄體及表觀遺傳等多層次分子變異驅動的高度異質性疾病。多體學資料整合提供深入解析腫瘤內複雜分子交互作用的機會，為精準腫瘤醫學開啟新的可能性。然而，現行方法多仰賴群體層級的分子相互作用網路，難以充分捕捉患者特異的分子變異；傳統深度學習方法亦常忽略不同體學模態間的語意對齊，導致其生物意涵與預測效能受限。本研究提出一種結合單樣本網路（Single-sample Networks, SINs）、圖注意力網路（Graph Attention Networks, GATs）及監督式圖對比學習（Supervised Contrastive Learning）的整合式跨體學計算框架，以強化精準醫療之多體學資料整合能力。具體而言，我們利用樣本特異性加權相關網路（Sample-specific Weighted Correlation Network, SWEET），為每位病患建構個體化的分子交互作用圖譜，以捕捉病患內部特有的分子異質性；接著，設計多視角GAT編碼器，產生節點層級（基因特異）與整體圖層級（患者特異）之嵌入表徵，完整保留局部與全域的分子資訊。此外，透過監督式圖對比學習，有效對齊不同體學模態的表徵並保持疾病表型間的區辨能力。我們在乳癌、肺腺癌及腎癌泛癌種資料集等多種癌症亞型分類任務中，展示優於現有方法，如MOGDx與MOGONET，分別達到91.01%、89.29%及98.73%的分類準確率。然而，本框架在癌症復發預測（卵巢癌及肝細胞癌）以及藥物反應預測（Carboplatin、Cisplatin、Fluorouracil）等較具臨床異質性的任務中，表現則相對有限，凸顯出此類預測任務的分子訊號更為複雜且微弱。模型可解釋性分析透過GNNExplainer顯示多個超越傳統亞型標記的重要基因與微小RNA，並挖掘轉錄體與表觀遺傳層次間之分子調控關係。路徑富集分析更進一步驗證細胞週期調控、細胞老化、蛋白聚醣訊號及細胞激素訊息傳遞等關鍵生物路徑的參與，顯示本模型預測結果之生物合理性與挖掘機制之潛力。本研究結果顯示結合單樣本網路、跨體學圖注意力機制與圖對比學習的跨體學分析計算框架，能夠捕捉癌症的分子特徵，提升多體學資料的預測效能與生物解釋能力，為精準腫瘤醫學提供新方向。;Cancer is marked by complex molecular heterogeneity driven by genomic, transcriptomic, and epigenomic alterations. Integrating multi-omics data provides unprecedented opportunities to interpret complicated molecular interactions, advancing precision oncology. However, conventional approaches, depending on population-level network analyses, fail to capture patient-specific molecular variation adequately. Additionally, traditional deep learning methods ignore semantic alignment across heterogeneous omics modalities, limiting their biological interpretability and predictive performance. In this study, we propose an integrative computational framework that combines single-sample networks (SINs), graph attention networks (GATs), and supervised contrastive learning to enhance multi-omics integration for precision medicine applications. Specifically, we utilize the sample-specific weighted correlation network (SWEET) approach to construct individualized molecular interaction graphs for each patient, capturing unique intra-sample molecular heterogeneity. Subsequently, we implement multi-view GAT encoders that generate fine-grained node-level (gene-wise) and comprehensive graph-level (patient-wise) embeddings, thus preserving local and global molecular contexts. To further refine feature discrimination and robustness, supervised contrastive learning is integrated, facilitating the alignment of omics-specific embeddings while maintaining clear distinctions among phenotypic classes. Extensive evaluations on diverse cancer cohorts, including breast invasive carcinoma (BRCA), lung adenocarcinoma (LUAD), and kidney cancer pan-cancer cohort (KIPAN), demonstrate that our framework exceeds state-of-the-art multi-omics methods such as MOGDx and MOGONET in subtype classification tasks, achieving accuracies of 91.01%, 89.29%, and 98.73%, respectively. However, predictive performance was comparatively limited in more clinically heterogeneous tasks such as recurrence prediction (ovarian cancer [OV] and liver hepatocellular carcinoma [LIHC]) and drug response modeling (Carboplatin, Cisplatin, Fluorouracil), highlighting intrinsic complexities and subtler molecular signals associated with these outcomes. Comprehensive interpretability analyses employing GNNExplainer identified biologically relevant genes and miRNAs beyond established subtype markers. These analyses revealed novel transcriptional and epigenetic regulatory interactions, with pathway enrichment studies validating involvement in critical biological processes, including cell cycle regulation, cellular senescence, proteoglycan-related signaling, and cytokine-mediated pathways. These findings highlight the biological possibility and potential mechanistic insights. In summary, this study demonstrates a computational framework leveraging single-sample networks, cross-omics graph attention mechanisms, and contrastive learning for precise molecular characterization in cancer. The proposed approach improves the predictive accuracy and interpretability of multi-omics integration, offering promise for precision oncology. |
